The Agent Skills Directory

[SAFE]: The skill integrates with official Prime Intellect infrastructure. All identified external resources (GitHub repositories, API endpoints, and documentation sites) belong to the vendor 'primeintellect-ai'.
Repositories: github.com/PrimeIntellect-ai/prime, github.com/PrimeIntellect-ai/verifiers, github.com/PrimeIntellect-ai/prime-rl.
Domains: primeintellect.ai, pinference.ai.
[SAFE]: The skill provides instructions for managing sensitive data (API keys, SSH keys, and sandbox secrets) using established security best practices such as environment variables and platform-level encryption.
[SAFE]: The 'Sandboxes' feature described in the documentation is specifically designed to isolate and execute potentially untrusted AI-generated code in disposable virtual machines, which is a significant security enhancement for AI agents.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an ingestion surface for untrusted data (AI-generated code and community RL environments).
Ingestion points: Data and code ingested via prime env install, prime env pull, and code executed via prime sandbox run or the Python SandboxClient.
Boundary markers: The skill relies on full virtual machine isolation (sandboxes) as the primary boundary between untrusted code and the host environment.
Capability inventory: The platform provides extensive capabilities including command execution, network access (tunnels), file system operations, and SSH access to remote pods.
Sanitization: Security is enforced through environment isolation and user-defined network allowlists/denylists for sandboxes.

prime-intellect